Optimal Scheduling of Heat-Integrated Multipurpose Plants - Industrial

Industrial & Engineering Chemistry Research 2014 53 (44), 17098-17111 ... Sequential Methodology for Scheduling of Heat-Integrated Batch Plants. Iskan...
2 downloads 0 Views 3MB Size
Ind. Eng. Chem. Res. 1994,33,3168-3186

3168

Optimal Scheduling of Heat-Integrated Multipurpose Plants Lazaros G. Papageorgiou, Nilay Shah, and Constantinos C. Pantelides’ Centre for Process Systems Engineering, Imperial College of Science, Technology and Medicine, London SW7 BBY, United Kingdom

A systematic mathematical framework for scheduling the operation of multipurpose batch/ semicontinuous plants involving heat-integrated unit operations is presented. The approach advocated takes direct account of the trade-offs between maximal exploitation of heat integration and other scheduling objectives and constraints. Both direct and indirect heat integration are considered. In the former case, heat transfer takes place directly between the fluids undergoing processing in the heat-integrated unit operations, and therefore a degree of time overlap of these operations must be ensured. It is shown that this involves only relatively minor modifications to existing detailed scheduling formulations. Indirect heat integration utilizes a heat transfer medium (HTM) which acts as a mechanism both for transferring heat from one operation to another and for storing energy over time. This provides a degree of decoupling with respect to the timing of the operations involved. The mathematical formulation presented in this paper is based on a detailed characterization of the variation of the mass and energy holdups of HTM over time. In particular, it takes account of the limitations on energy storage due to heat losses to the environment. A modified branch-and-bound procedure is proposed for the solution of the resulting nonconvex mixed integer nonlinear programming problem. 1. Introduction

Recent work (Vaselenak et al., 1986;Lovell-Smith and Baldwin, 1988) has highlighted the need for efficient utilization of energy in the operation of batch and semicontinuous plants. However, in contrast to the extensive amount of work already published on energy integration in continuous plants, relatively little has been reported in the literature on this aspect of the operation and design of flexible multipurpose plants. Vaselenak et al. (1986) consider the problem of determining the best pairing between batch vessels containing hot fluids that require cooling and cold fluids that require heating over a given time period. They examine two cases: in the first, the fluids return to the original holding vessels after heat exchange, while in the second, the fluids remain entirely within the holding vessels while heating and cooling media accomplish the exchange. They present a heuristic procedure to determine best pairings when final temperatures are not limiting, and a mixed integer linear programming (MILP)formulation for the case when they are. A given schedule of operations is implicit in the problem description. Ivanov e t al. (19921, and Peneva et al. (1992) consider the problem of designing a minimum total cost heat exchanger network for given pairwise matches of batch vessels. They examine the cases where heat exchange takes place directly between the two fluids (Ivanov et al., 1992), and also where intermediate agents are used (Peneva et al., 1992). The majority of all other papers published in the literature are based on the concept of pinch (Linnhoff e t al., 1982), modified in order to accommodate the complications introduced by the time-varying operation of batch processes. The most rudimentary approach (Linnhoff e t al., 1988) treats batch plants as pseudocontinuous ones, time-averaging the energy requirements of the various processing steps and applying a cascade analysis to the grand composite curve that arises from the pseudocontinuous plant. This “time

* Author t o whom correspondence should be addressed.

average model” (TAM)approach is useful in that it sets rigorous targets for minimum utility consumption, against which actual consumption can be compared. Kemp and MacDonald (1987, 1988) do take into account the time-varying nature of batch processes by assuming a fixed production schedule. Opportunities for heat exchange are identified by inspecting the energy requirements and temperatures of the processing steps. They utilize a “time-temperature cascade” ~TTC) approach, where both direct and indirect heat exchange are possible. The use of heat storage in the latter case means that heat flows can be both time and temperature cascaded. The authors assume that direct heat transfer is always preferable to indirect heat transfer, and are thus able to derive targets for both energy storage and utility consumption. However, the utility consumption targets are over-conservative as no rescheduling is undertaken, and tend to be very different from the minimum targets predicted by the TAM. Obeng and Ashton (1988) consider a fured sequence of operations and identify heat exchange opportunities using a time-slice model (TSM) which determines the intervals over which only direct heat exchange may occur, but otherwise bears some resemblance to the TTC model. This is used not only to predict utility consumption targets, but also to identify opportunities for adjusting the timing of operations to reduce these targets. An example is presented to demonstrate that, with appropriate retiming of operations, these targets approach the absolute ones predicted by the TAM. Kemp and Deakin (1989)extend the work of Obeng and Ashton by considering both the direct and indirect modes of heat exchange as well as retiming of operations. More recently, Corominas e t al. (1993) considered the problem of designing a minimum cost heat exchanger network and a heat exchange strategy for multiproduct plants operating in campaign mode. A prespecified campaign comprising product batches and hot and cold stream information is examined to determine the best hothold stream match according to rules aiming to achieve maximum heat exchange. Since the authors only consider direct heat transfer, it may be necessary

0888-588519412633-3l68$04.50/0 0 1994 American Chemical Society

Ind. Eng. Chem. Res., Vol. 33, No. 12, 1994 3169

to retime the operations slightly to ensure that the paired streams are available simultaneously. The matches selected imply the heat exchanger network design, thus allowing the campaign to be costed. Different campaigns may be proposed, costed, and compared using this approach. Despite its clear importance, the minimization of the cost of external utilities consumed is not usually the primary objective in scheduling the operation of a multipurpose plant. This is a consequence partly of the paramount demand for timely satisfaction of the multiple production requirements imposed on these plants, and partly of the often small proportion of energy costs compared t o the high value of the raw materials and products produced in many such plants ( e g . , in the pharmaceutical industry). It could, therefore, be argued that optimizing the exploitation of any heat integration opportunities afforded by a fzxed production schedule that already achieves all other plant objectives is indeed a reasonable approach. However, in general, even optimal production schedules tend to be quite degenerate, in the sense that there often exist a large number of different schedules, all of which can achieve a given set of production requirements. However, the potential for heat integration could vary significantly from one such schedule to another. Furthermore, in some industrial sectors ( e g . , food, dairy, brewing) that employ multipurpose plants, energy costs do form a significant proportion of the total production cost, and thus have t o be balanced properly against other costs, such as those of the raw materials and manpower, and the value of the products (see, e g . , Mignon and Hermia, 1993). On the basis of the above discussion, we argue that heat integration must be considered as an integral part of the problem of scheduling the production in a given plant, with the cost of utilities incorporated within the overall economic objective of maximizing the net value of the production over a given time horizon. We consider two potential modes of heat integration. In direct heat integration, heat exchange occurs between two processing steps taking place simultaneously in two separate items of equipment coupled by a heat exchanger. On the other hand, indirect heat integration makes use of a heat transfer fluid that can also be used to store energy. This provides additional operational flexibility as it removes, to a certain extent, the need for synchronization of the various processing steps involved. However, because of heat losses to the environment, the periods over which energy can be stored are often of limited duration. This paper proposes a systematic mathematical framework for the exploitation of heat integration in batch plqnt operation. In particular, it demonstrates how heat integration aspects can be incorporated within a general mathematical formulation for the scheduling of multipurpose plant operation (Kondili et al., 19931, thus ensuring that all the relevant scheduling aspects are taken into account. More specifically, we consider a short-term scheduling problem that seeks to determine the optimal utilization of the available plant resources (processing equipment, storage capacity, utilities, manpower, etc.) over a given time horizon. The plant may manufacture one or more products, and minimum and/ or maximum production requirements may be imposed on them a t various times during or a t the end of the horizon. Assuming that the problem posed is feasible, the objective is to select the schedule that maximizes

an economic performance measure taking into account the value of the products and the cost of raw materials and utilities. The rest of this paper is organized as follows. The next section considers the case of direct heat integration. It provides a brief review of the Kondili et al. formulation, and introduces extensions which enable it to accommodate the complications arising from direct heat integration. An example of a simple heat-integrated process is used to illustrate the approach. Section 3 analyzes the indirect heat integration case and examines methods for the solution of the resulting scheduling problem. Finally, some concluding remarks on the strengths and limitations of the overall approach are made in section 4.

2. Direct Heat Integration 2.1. Problem Characteristics. We assume that the processing equipment available in the plant includes a t least one pair of units ( e g . , a reactor and a batch still) which are coupled to each other through a heat exchanger. This permits heat exchange to take place between two processing steps ( e g . ,an exothermic reaction and a distillation operation) that are performed simultaneously in these units. We further assume that the precise operating mode and characteristics of two heat-integrated processing steps taking place in a given pair of equipment items are known a priori. This information includes the processing times of each of the two steps, as well as a fixed offset between their starting times. Such an offset may be necessary in order to maximize the potential for heat integration. For instance, in the case of heatintegrated reactioddistillation operations, it may be appropriate to start the reaction step earlier than the distillation step. This would allow sufficient time for the reaction t o reach the operating temperature a t which the heat consumed by the distillation is to be generated. Papageorgiou et al. (1994) describe a procedure for determining optimal policies for thermally coupled batch operations using dynamic optimization techniques. In general, the instantaneous rates of heat production and consumption of the two coupled tasks may be different in magnitude. We therefore assume that the equipment is also fitted with heat exchange facilities that allow the use of external utilities ( e g . , steam, cooling water) to supplement the heating and cooling loads provided through heat integration. The existence of these additional facilities also makes possible the use of each of the equipment items in a pair for carrying out the corresponding processing step separately, if necessary, i.e. without the need for heat integration. Of course, the processing times of the individual steps when performed separately may well be different from those for the same steps when heat integration is employed. In any case, the instantaneous rates of consumption of external utilities over the duration of each step are assumed to be k n o w n . 2.2. Review of the Kondili et al. (1993)Formulation. The mathematical formulations presented in this paper are extensions of the general formulation of the short-term production scheduling problem presented by Kondili et al. (1993). A key feature of this approach is the use of the state-task network (STN) process representation (Figure la). STNs are directed graphs comprising two types of nodes. State nodes (denoted by circles) correspond to materials of different types,

3170 Ind. Eng. Chem. Res., Vol. 33, No. 12, 1994

Reaction

Distillolion

PrOdUCt2

FeedB

(a) State-Task Network Process Representation

H

-m

1

2

3

4

.........

,,e



-c

.’

..............

H

H+1

(b) Uniform Time Discretization Figure 1. Key concepts of Kondili et al. (1993) formulation.

including feeds, intermediates and final products, while task nodes (denoted by rectangles) represent physical, chemical, or biological transformations of these materials. One advantage of the Kondili et al. approach is its ability to accommodate complex processes with material recycles, shared intermediates, and multiple processing routes t o the same type of material. It also allows mixing and splitting of individual batches, as well as significant flexibility in the utilization of multipurpose processing equipment. The mathematical formulation is based on the discretization of the time horizon into a number of intervals of equal duration (see Figure lb). System events (task starts and finishes, changes in resource availability, product demands, etc.) are allowed to occur only a t the boundaries of time intervals. The key variables of the formulation characterize the utilization of resources over the time horizon, as follows: I. Utilization of processing equipment. (a) Equipment allocation variables. Wqt = 1if equipment unit j starts processing a batch of task i at time t; 0 otherwise. (b) Batch size variables. Bqt = amount of material that starts undergoing task i in u n i t j a t time t. 11. Utilization of dedicated storage capacity. S,t = amount of material in state s held in storage over time interval t. 111. Utilization of utilities and manpower. Uut = amount of utility u being used over time interval t. The above variables are subject to a number of constraints, including (a) processing equipment allocation constraints expressing the fact that multipurpose equipment can carry out at most one task over any given time interval, (b) processing and storage equipment capacity constraints, (c) utility utilization and availability constraints, and (d) mass balance constraints describing the variation of the inventory of each material over time. The precise mathematical form of these constraints is not important for the purposes of this paper, and the interested reader is referred t o the Kondili et al. paper for more information.

Unit j

I

Unit j’

-e-’

Task i

i Task i’

I

Time Figure 2. Triplet {(i, j ) , (i’,j’), e}.

Finally, the objective function is a measure of the economic performance of the plant over the available time horizon, including the value of the products, the cost of the raw materials and the cost of utilities. In a somewhat simplified form, it can be written as H

The difference (S,&+1 - S,,O)in the first term reflects the net increase in the inventory of material s over the time horizon, and this is multiplied by the unit value of this material, C,. The second term represents the cost of the utilities used by the plant over the horizon, with Cut being the unit cost of utility u over interval t. The inclusion of utility costs in the objective function (1) makes this formulation suitable for considering opportunities for heat integration on a firm economic basis that takes direct account of the trade-offs between energy savings on one hand, and increased production and satisfaction of scheduling constraints (such as timely delivery of orders) on the other. 2.3. Mathematical Formulation for the Direct Heat Integration Case. The additional complications introduced by direct heat integration can be accommodated through a few simple extensions of the basic scheduling formulation. Each of the allowed heat exchange matches is represented by a “triplet” comprising two pairs of tasks and units, and an offset with respect to their starting times, as illustrated in Figure 2. Thus, the triplet {(ij),(i‘j’), 0) represents the possibility of exchanging heat between task i when

Ind. Eng. Chem. Res., Vol. 33, No. 12, 1994 3171 e----

-

S&d-alone Opaation

-- --

Hear Integrated Opaations

Stand-alone Operation e

Figure 3. STN modification for modeling heat-integrated operations.

performed in unitj, and task i’ when performed in unit j , with a time offset of 8 time intervals between their corresponding starting times. A given pair of processing

units G , j ) coupled by a heat exchanger may be used for canying out more than one pair of thermally-coupled processing steps, but each such possibility will be represented by a separate triplet. The first issue that needs to be addressed is allowing for the possibility of each of the operations (i,i’> actually taking place without any heat integration by relying exclusively on the use of external utilities. To achieve this, we split each of the original STN tasks i , i’ into two, one corresponding to the heat-integrated operation of the task and one t o the stand-alone operation, as shown in Figure 3. Here the notation ilj is used to denote task i performed specifically in vesselj with heat integration. It should be noted that both the standalone tasks and the heat-integrated ones involve the same transformations of material. The second issue to be addressed is that of ensuring that the heat-integrated tasks ilj and i’lj’ always start with the specified time offset 8. This can be achieved by introducing constraints relating the corresponding equipment allocation variables in the scheduling formulation. In particular, we need to ensure that variables Wiljjt and Wcy,y,t+ehave the same value; z.e., a batch of task ilj starts in unitj at time t if and only if a starts in unitj’ a t time t 8: batch of task

i‘v

+

Alternatively, the number of binary variables and constraints can be reduced by using (2) to eliminate half of the binary variables involved, thus introducing only one binary variable for each triplet { ( i ,j ) , (z’, j ) ,e} a t each time instance t. Finally, the ratio of the batch sizes of the heatintegrated tasks will usually have to lie within given limits in order for their combined operation to be feasible. This requirement can be expressed mathematically as

where pg? and pm? represent lower and upper bounds on the rad0 of t h l t a t c h sizes of the two heat-integrated tasks. 2.4. Example of Direct Heat Integration. We consider a plant manufacturing two products, Productl

and Product2, according to the following recipe which corresponds to the STN of Figure la. Reaction: Allow a mixture of 60% of FeedA and 40% of FeedB to react for 2 h in order to produce intermediate ReactProd. This is an exothermic reaction requiring cooling throughout its duration. For a batch size of B metric tons (t),a constant cooling water flow rate of 1.59 0.1OB t’h is needed throughout the reaction. Filtering: Filter intermediate ReactProd for 1 h to form intermediate Filterprod. A small amount of solid waste is also produced at this step, but this can be ignored for the purposes of this example. Distillation: Distill intermediate Filterprod t o produce 75% of Productl and 25% of Product2 after 2 h. For a batch size of B t, a constant steam supply to the reboiler of 0.044 0.0035B t’h is required throughout the operation. The available processing equipment comprises a 60 t reactor, a 80 t filter, and a 70 t batch distillation column. The minimum allowable size of batches being processed in the reactor and the column cannot be less than 25% of their normal capacity. The corresponding figure for the filter is 10%. Sufficient dedicated storage is provided for all raw materials and final products, while storage vessels of 100 t are available for each of the two intermediates. Unlimited availability of steam and cooling water is assumed, with corresponding unit costs of 200 and 4 relative cost units (rcu) per metric ton, respectively. The time horizon of interest is 48 h. The unit values for the two products are assumed to be equal at 5 rcdt. The objective function is the maximization of the combined value of the production over the time horizon, incorporating both the value of the products produced and the cost of the external utilities (steam and cooling water) consumed. Assuming appropriate temperature levels, there is in principle the opportunity of exchanging heat between the Reaction task which requires cooling, and the Distillation task which requires heating. An equipment configuration that would permit this heat exchange to be realized is illustrated in Figure 4. More specifically, this heat exchange match would involve task Reaction being performed in unit Reactor and the Distillation task in unit Column with an offset of 1h with respect t o their starting times. Due to lower temperature differences, the duration of the reaction when operated in heat-integrated mode is increased from 2 t o 3 h.

+

+

3172 Ind. Eng. Chem. Res., Vol. 33,No. 12, 1994

Figure 4. Direct heat exchange network for example process. 2hr

Zhr

3hr

Figure 5. Modified STN for example process.

As expected, the demands posed by these tasks on external utilities are also modified. Thus, for a batch size of B t, the heat-integrated reaction task requires a cooling water flow rate of 1.0 0.06B tk during the first hour of its operation only; thereafter all its needs are satisfied through heat exchange with the contents of the distillation column reboiler. On the other hand, the heat-integrated distillation task requires a constant steam flow rate of 0.020 0.0016B t/h throughout its duration supplementing the energy received from the heat exchange with the reactor contents. The STN for the process taking into account both stand-alone and heat-integrated operations is shown in Figure 5. The optimal schedule for the case without heat integration is shown in Figure 6. The number above each line segment in the Gantt chart denotes the task being carried out (see key in right hand side margin of chart). The number below the segment is the amount of material being processed (in metric tons). The amounts of the two products produced are 945 and 315 t, respectively, while the costs of steam and cooling water consumption are 2080.8and 1275.1rcu. Overall, this results in an objective function value of 2944.1 rcu. Figure 7 shows the optimal schedule for the case where heat integration is allowed. It can be seen that this involves both stand-alone operations and integrated ones. In particular, a batch of the stand-alone Reaction (task 1) has to be performed a t the start of the horizon, a period during which the integrated operation (tasks 4 and 5) cannot take place due to the unavailability of intermediate Filterprod feeding the Distillation task. More interestingly, stand-alone reaction tasks are also performed sporadically throughout the horizon. This

+

+

feature of the optimal solution is due to the complex trade-off between increased processing time and reduced utility consumption for the reaction task. On the other hand, since heat integration leaves the processing time of the distillation unaltered, no stand-alone distillation operations take place. The amounts of the two products produced in this second case are 720 and 240 t, respectively, which are lower than the corresponding figures for the case without heat integration. This is to be expected as the heat integrated reaction task takes longer than its stand-alone counterpart. On the other hand, the costs of the utilities consumed are reduced to an even larger extent, down to 726.4 rcu of steam and 429.0 rcu of cooling water. Overall, this results in an objective function value of 3644.6 rcu, representing a 23.8% improvement over the case without heat integration. The mathematical formulations of the two cases as MILPs involve 142 and 188 binary variables, respectively. The optimal solutions were obtained after the examination of 17 and 5 nodes of the search tree, respectively. A 5% margin of optimality was used for the branch-and-bound procedure. The optimal solutions were respectively within 4.76% and 0.37% of the objective function values of the corresponding fully relaxed linear programmes. 3. Indirect Heat Integration 3.1. Problem Characteristics. The indirect mode of heat exchange involves the use of a heat transfer medium (HTM) which is assumed to be generally available a t a number of different temperature levels,

Ind. Eng. Chem. Res., Vol. 33. No. 12,1994 3173

unit

TUkl

NO HEAT INTEGRATION

..lllucnoN

.. m m o

.. D U l l l l A l I O

(a) Utilization of processing equipment @OS w LOA (I1Jwm 1993) 8 h p i d Cdbp 1993 ImprldCoUe* s l32%17 m TbuMr241994

R e s

NO HEAT INTEGRATION

SIBAM

I1

................ ........... .................. ......... ..... ....................... ............... .. ..........................., ... ..............................

................................................... .......................

0

m

(b) Utilization of external utilities Figure 6. Optimal schedule without heat integration.

each kept in a separate, well-mixed, and possibly insulated storage tank. The existence of storage capacity for the HTM effectively removes the requirement for strict synchronization between the tasks producing and those consuming heat, thus permitting more flexibility in plant operation. 3.1.1. Network Connectivity. As shown schematically in Figure 8, it is assumed that direct pipe connections exist between the processing units on one hand and the HTM storage tanks on the other. This allows each unit to be used independently of all others in the network, provided of course that sufficient amounts of HTM are available at the required temperature(s). Each processing unit may be connected t o more than one HTM storage tank. However, we assume that only one of its inlet and one of the outlet connections are utilized during any one heat-integrated task taking place in a given unit. As for the direct heat integration mode, it is also assumed that the processing units in the heat transfer network are equipped with facilities allowing the use

of external utilities, either supplementing or replacing the use of the HTM. 3.1.2. Operating Policies for Heat-Integrated Operations. The operating policy for a batch operation can be expressed in terms of its duration and the variation of the available control variables (e&., the reflux ratio in a batch distillation column, or the rate of steam supply to its reboiler) over this duration. For heat-integrated batch operations, the heat load provided by the heat integration can be viewed as an additional manipulated variable. The above operating decision variables may be specified in an ad hoc manner aiming t o satisfy various endpoint constraints (eg., purity of the final products) or path constraints ( e g . ,maximum temperature throughout the operation). Alternatively, they may be chosen so as to optimize some performance criterion ( e g . , maximize yield) subject to the same constraints. This usually involves the solution of complex dynamic optimization problems. For the purposes of this paper, we assume that the operating policy for each operation (heat-integrated or

3174 Ind. Eng. Chem. Res., Vol. 33, No. 12, 1994

DIRECT HEAT INTEGRATION

Unit

(a) Utilization of processing equipment #BSS VL LOA (17 Jum 15W) 0 Imparid C d b p 1993 Impridc&p u 154:n m TbuMUUl994

RlUmiLl#U

DIRECT HEAT INTEGRATION

n

1 '

c

.................., ..................... .... .. .... ...........,.. ........,.. .... ..,.......... .... ... . ...... .............. ............,........... ....................................... ...................... .. ......... . .. .

y

(b) Utilization of external utilities Figure 7. Optimal schedule with direct heat integration.

otherwise) is fixed and k n o w n a priori. We further assume that the time profiles of external utility consumption and the HTM heat load in each heatintegrated unit operation are k n o w n piecewise constant functions of time over the task duration (see Figure Qa). These functions may be independent of the batch size, or vary linearly with it. For instance, the heat required for preheating a metal reactor vessel before loading the reactants is effectively determined by the mass of the vessel. On the other hand, the cooling or heating load demanded by the reactor contents during the course of the reaction is essentially proportional to the amount of material undergoing processing. Once the duration and operating policy for a heatintegrated operation are fixed, the time variation of the temperature of the material undergoing the operation can also be determined, either by simulation or experimentally. In general, this variation will be a k n o w n continuous function of time, of the form shown in Figure 9b. The overall plant schedule must ensure that any heatintegrated operation taking place receives both the

required external utilities and the necessary heat load from the HTM at the correct levels throughout its duration. This is not entirely straightforward because the temperature of the stored HTM may vary over the time horizon of interest, which in turn affects the temperature difference that drives the heat transfer between the unit operation and the HTM flowing through it. Thus we need to determine suitable time profiles for the flow rates of the HTM. As will be explained later, in general these may be discontinuous functions of time over the duration of the task (see Figure 94. 3.1.3. HTM Storage. Unless the HTM storage vessels are heavily insulated and/or the corresponding HTM temperatures are very close to ambient levels, some heat losses (or gains) to (or from) the environment will be unavoidable. Depending on the thermal conductivity of the material of construction of the storage vessel, and also the geometry of the latter, the effective heat transfer area for these interactions with the environment may be constant or it may depend on the amount of material held in the tank at any given time.

Ind. Eng. Chem. Res., Vol. 33, No. 12, 1994 3176

Figure 8. Heat transfer network for indirect heat integration. ,_...... .................................................... w.w~.I

i-j

l..l ...I

.... .... ___,

....

...

(

260.mi..

:r.......

,-I

240.00 :..

435.00 430.00

425.00;...

220.00Y ' '

..< ...I........... i ............ :........... :................

200.00;.

,

0.00

1.00

C---

2.00

3.00

I

4.00

Task Duration

(a) External Utility Consumption and HTh4 Heat Loads for Heat-Integrated Operations

I#\

,... . :..I: .......... 1............ :...........;...............

420.00L.

-

0.00

1.00

2.00

3.00

4.00

TaskDuration

(b) Temperaturein Heat-Integrated Operations

-

TaskDuration

-

(c) HThl Flowrate through Heat-Integrated

Operations

Figure 9. Characteristics of heat-integratedoperations.

The other important factor in determining the rate of heat loss or gain is the ambient temperature, which may vary over the time horizon of interest. We assume that this variation is known a priori, and that the overall coefficient for heat transfer between each storage tank and the environment is constant and known. Because of the heat losses or gains mentioned above, and also due to differences in the actual temperature at which HTM is produced by different tasks, the temperature of the HTM in each tank will generally vary over the time horizon of interest. However, in the interests of proper operation of the heat-integrated tasks, it is usually desirable that these temperatures be maintained within given upper and lower bounds. In practice, the main mechanism for achieving this is the adjustment of the timings of the various tasks taking place over the horizon as determined by the operating schedule, and the flow rates of the HTM used by each task. 3.2. Mathematical Formulation for the Indirect Heat Integration Case. The operation of the heat integration network during the time horizon of interest can be described in terms of the following functions of time: S,(t) = amount of material in HTM storage tank s at time t

of material in HTM storage tank s a t time t q&) = heat load t o the HTM from contents of unit j carrying out heat-integrated task i a t time t fij(t) = flow rate of HTM through unit j carrying out heat-integrated task i a t time t @rt(t) = temperature of HTM leaving unit j carrying out heat-integrated task i at time t The approach we follow is to perform rigorous mass and energy balances on each HTM storage tank and each heat-integrated operation. These lead to differential equations in time. We then integrate these equations over each discrete time interval, using trapezoidal approximations to the various integrals arising from the time integration. The resulting discrete time formulation is expressed in terms of finite-dimensional sets of parameters corresponding t o the values of the above variables at the time interval boundaries. As shown later in this section, the flow rates f-,.(t) and the temperatures are not necessarily continuous a t the interval boundaries. We therefore need t o introduce separate parameters representing the values of these variables at the end of one interval and the start of the next. @ ( t ) = temperature

@rt(t)

3176 Ind. Eng. Chem. Res., Vol. 33, No. 12, 1994

I

Taski, in unit j,

Task L in unit j

I Task i, in unit j,

Task i, in unit j,

Ki2 = (Jd Figure 10. Mass balance on HTM storage tank s.

The discretization parameters to be determined by the scheduling algorithm are therefore the following: SSt= amount of material in HTM storage tank s a t time point t OSt= temperature of material in HTM storage tank s a t time point t qgt = heat load to the HTM from contents of unit j carrying out heat-integrated task i over time interval C

f i j , = flow rate

of HTM through unitj carrying out heatintegrated task i a t the start of time interval t = flow rate of HTM through unit j carrying out heat-integrated task i a t the end of time interval t Ogl;lt = temperature of HTM leaving unit j carrying out &eat-integrated task i at the start of time interval t O'pl;lt = temperature of HTM leaving unit j carrying out ieat-integrated task i a t the end of time interval t In addition to the above, the formulation also involves the variables of the original scheduling formulation described in section 2.2. 3.2.1. STN Modification. As in the case of direct heat integration, we start by modifying the process STN to take account of the possibility of the same task taking place either in a stand-alone or in a heat-integrated mode. This can be achieved by splitting each original task into two, both involving the same transformations of material, but otherwise being different with respect to their processing times and/or consumption of external utilities (cf. Figures 3 and 5). 3.2.2. Mass Balances on "I'MStorage Tanks. As illustrated in Figure 10, HTM in ssorage tank s is produced and consumed by subsets T, and T, respectively of the heat-integrated tasks in the network. In general, each task i may take place in more than one processing unit j at the same time. A mass balance in the continuous time domain yields the differential equation

djt

where Ki is the set of processing units that are suitable for task i. By integrating (4) over time interval t and using the trapezoidal rule to approximate the integrals of the flow rates, we obtain

W Figure 11. Energy balance on heat-integrated operations.

where 6 is the length of the discretization interval. 3.2.3. Energy Balances on Heat-IntegratedOperations. We now consider the energy balance on a heat-integrated operation i taking place in a unitj. This receives HTM from a storage tank, denoted by sg, and discharges it into another HTM storage tank, denoted by Sg (see Figure 11). The temperatures of the HTM entering and leaving the unit at time t are denoted by O,,(t) and OPut(t), while the temperature of the fluid surrounding the heat transfer jacket (or coil) is assumed to be uniform and is denoted by TgW. We assume that the dynamic behavior of the jacket is fast compared to the rest of the operation, and that the rate of heat transfer qg(t) between the fluid in the unit and the HTM flowing in the jacket can be described adequately in terms of an arithmetic mean temperature difference:

where Ug and Ag are the relevant heat transfer coefficient and area, respectively. An enthalpy balance on the HTM yields

where h,, and hi are the specific enthalpies of the HTM in the supply storage tank and the HTM leaving the heat-integrated operation respectively. Assuming a constant specific heat capacity, cp, for the HTM, we can write (6) as

From ( 5 ) and (6'1, we can eliminate the outlet temperature Oyt obtaining an expression for the flow rate fg in terms of the quantities qij, Tg, and Os,. Since the heat load q g is assumed to be a piecewise constant function of time while the temperatures T i and Os..are continuous functions, we can deduce that fij(t)wifl, in general, be a discontinuous function of time, of the form shown in Figure 9c. The temperature O r t ( t )will be a similarly discontinuous function. The temperature of the fluid surrounding the jacket a t discrete time points 0 (0 = 0, ...,p i ) relative to the start of heat-integrated task i taking place in-unit j is also assumed t o be known, and denoted by Tge. The corresponding temperature in absolute time, Tit will depend on when the task actually started. This can be expressed as

Ind. Eng. Chem. Res., Vol. 33, No. 12,1994 3177 .!:--'.:.~%y$?$p . ,. . ..r.i;?i 8L.i-l: ti

Task i starting in unit j at tim t-2:

=1

-.

r-I:

=1

WI-1

K=1

1:

I

I

I

I

I

I

I

t-3

1-2

I-!

I

r+1

1+2

t+3

I

I

I

I

I

)

Time Axis

Figure 12. Heat load contributions over interval t .

where pi is the duration of task i. I t should be noted that, because of the allocation constraints in the scheduling formulation (Kondili et al., 1993), a t most one of the binary Wg,t-s variables in the summation on the right hand side can be 1. For instance, if the operation i started in unit j at time t - 2, then Wij,t-~= 1, and, assuming that pi > 2, (7)will reduce correctly to Tot.= T ~ z It. may also be worth mentioning that (7)implies that Tot = 0 if operation i is not taking place in unit j a t time t . This is acceptable as the function T&) is, in fact, undefined at such times. We can now produce discrete time versions of constraints ( 5 ) and (6') by writing them a t the start and end of each interval and using (7) to eliminate the quantity Tof:

fv~p(@Tt

(8) t = - @8y.t+1) V i ,j E Ki, The final task is to relate the heat load qCt to the amount of material undergoing the heat-integrated operation. Over time interval e relative to the start of the operation, this heat load is given by

g,, + y;$3

i3 = 0, ...,pi- 1

4..= 91

a=o

U&I8(@,- 7-b)

vs

(11)

where u.W and h,(t)are respectively the specific internal energy and enthalpy of the contents of the tank, TBmb(t)is the ambient temperature, and Usand A, are respectively the heat transfer coefficient and the effective heat transfer area for heat losses to the environment. It should be noted that A, may vary over time as it may depend on the amount of material held in the tank. By combining (11) with (6), we can eliminate the specific enthalpy of the streams entering the tank, h&, to obtain

By integrating the above over time interval t , taking account of the fact that q&) is constant during the interval, and replacing all integrals on the right hand side by trapezoidal approximations, we obtain the discrete time relation:

(9)

where v!fi and v? . ,." are -given constants and B is the batch si;. When considerine an interval t in absolute time, the heat transfer rate :i a given ( i ,j ) pair may be attributed to a batch of task i that started in u n i t j a t time t , or to batches that started a t earlier times t - 0, i3 = 1, ...,pi - 1, and are still being processed (see Figure 12). To take this into account, we have to sum (9) over all possible 8, yielding Pi-1

is necessary to ensure that no contribution to the summation is made if task i has not started in u n i t j a t time t - 0 (corresponding to Wij,*-e= 0). 3.2.4. Energy Balances on IlTM Storage Tanks. An energy balance on the HTM storage tank s in the continuous time domain can be written (see Figure 13) as

(y!!'w..y.t-8 + Y(2)~ ~ B ~v i~, j -E Ki, ~ t) (10)

The multiplication of y t j by the binary variable WV,~-O

qij)-6 U @ , t

Us6

cc

f i j ~+ t

kT. jeK(

6jA,t+i-

- Tmb) + As.t+i(@s.t+i- T?:)

vs,t

2 (12)

For most HTMs of practical interest, the difference between internal energy and enthalpy is negligible. By W h e r assuming constant specific heat capacity, we can rewrite (12) entirely in terms of temperatures, as

3178 Ind. Eng. Chem. Res., Vol. 33, No. 12, 1994

Task i3 in unit j,

Task i, in unit j,

Energy losses

U, A, (0, - Tam? Figure 13. Energy balance on HTM storage tank 8 .

follows:

and

0,”” C e,,5

e:= v s, t (15) There may also be a n upper limit, q-, to the flow fgtest

+ AjtQ,t+1

rate of HTM that can be pumped through any given unit j. We therefore have

-

2

cmb) + A8,f+l(@s,f+l - cy:)

U,6 A,,(@,,-

2

CP

0 c fut vs,t

(12’) We assume that, for a given vessels, the area ktis a simple linear function of the holdup S,, of the form

- S*,% =

Ss.f+1%,,+1

er

+ t .Ut o

SU.I+l

2

ieT. jeK,

e

+-)qu,

fu,%

CP

+ #,Q,,f+l

-

2

k T , j6K8

U,6 [(a:) + d2)S8t)(08f -qmb,

“’1

1 2 (e+ ~z)s8,f+1)(0s,f+1 CP

2

v s, t (12”)

3.2.5. Capacity and Operability Constraints. The amount of HTM stored in storage tank s cannot exceed the tank capacity C,. Also, its temperature may have to be maintained within given lower and upper limits C 3 p and These restrictions lead to the bounds constraints

05

s,, 5 c,

v s, t

(14)

q”

V i,j E

K,,t

(16)

3.3. Solution of the Scheduling Problem. The scheduling problem that must be solved to determine the optimal operation of the heat-integrated plant is summarized below: Problem

P1

41)+ 4%,, (13) where the coefficients 4’)and d*) depend on the vessel

geometry and its material of construction. By using this expression in (12‘). we obtain the final form of the energy balance:

5

subject to

Ind. Eng. Chem. Res., Vol. 33, No. 12,1994 3179

2

OIS,,IC,

vs,t

(14)

and min 0,

orax

IOs,I

vs,t

(15)

Lower and upper bounds on the newly introduced variables may be established using appropriate bounds on the variables in the corresponding bilinear terms(Mccormick, 1976). This leads to constraints of the form

Tax

(16) v i , j E Ki,t 0 Ifw, I and all other constraints in the formulation of Kondili et al. (1993). The above mathematical optimization problem is a nonconvex mixed integer nonlinear program (MINLP). In particular, the energy balance constraints (8) and (12") introduce nonlinearities due t o bilinear terms of five different forms, namely fw,OZt, fwtOfgt,SstOst, f&t, and &Os,. In order to circumvent the problems posed by nonlinearity and nonconvexity, we consider an approximate linearization of the bilinear terms. Specifically, we define the following new variables:

FO,, -

fij,O$

V i , j E Ki, t

(17a)

FO:. Ut = f.. wt@'?ut wt

'# i , j E Ki, t

(1%)

-

FO,,

fijtO,,

V s, i

-

FO'Gsst = f..0s,t+l and

E

T,,j E Ki, t

b' 5, i E

(17c)

T,,jE Ki, t (17d)

-

SO,,= S,,O,, v s, t (174 which allows us to rewrite the energy balance equations (8) and (12") in linear form: The lower and upper bounds O y and Orax on the temperatures OStand the upper bound C, on the HTM holdups SStare given as part of the problem specification. However, appropriate upper bounds and for the flow rates fwt and fwt at the start and end of each time interval must be established. Similarly, we need lower bounds Oyt9"" and OfTti- and upper on the HTM exit tembounds OytPmax and peratures Ogt and Of$ at the start and end of the time intervals. The denvation of these bounds from the energy balance equations ( 5 ) and (6') and pumping rate constraint (16)is presented in the Appendix.

fr" ,

_.

-

-

= cp(FOtit-FO;ssiit)

V Z,j E Ki, t

(8')

3180 Ind. Eng. Chem. Res., Vol. 33, No. 12, 1994 The modified linear energy balance equations (8')and (12"') together with the linear bounds (18)can be used to replace the nonlinear energy balances (8) and (12") in the scheduling formulation of problem P1. Hereafter we refer to the MILP resulting from this substitution as problem P2. Because of its linear nature, problem P2 can, at least in principle, always be solved using an appropriate method such as branch-and-bound. On the other hand, it should be noted that P2 is only a relaxation of P1 and therefore the temperature, flow rate, and holdup profiles obtained by solving it may not satisfy the original energy balance equations. Also the objective function at the optimal solution of the MILP will be only an upper bound on the optimal objective function that can be attained by the system. In view of the above discussion, we must consider modifications to the standard branch-and-bound algorithm to be applied to this problem. In particular, we note that each integer-feasible node visited during the search for the solution of P2 corresponds to a feasible allocation of the processing equipment as determined by the corresponding values of the binary variables WG~. However, due to the relaxation of the bilinearities, the corresponding value of the objective function is generally an upper bound to the true value of the objective function a t this node. Now consider solving the restricted nonlinear programming (NLP) problem obtained from P1 by f ~ n g the binary Wijt variables a t the values that they have at an integer-feasible node. This corresponds to the solution of a nonconvex NLP, a feasible solution to which is also a feasible solution to P1. However, because of the nonconvexity, standard NLP algorithms are guaranteed to obtain only a local solution. This therefore provides a lower bound on the true objective function value for this node. In summary, at each integer-feasible node of the branch-and-bound search for the solution of P2, we establish two bounds on the corresponding objective function value of P1: an upper one from P2 and a lower one from the (restricted) P1. If the search were to examine all integer-feasible solutions of P2, then the maximum lower bound and the maximum upper bound encountered during this search would also bound the optimal solution of P1. However, if at any point during the search the upper bound to the objective function at the node being examined is smaller than the best lower bound established so far, then it is clearly not necessary to consider further this node or any of its descendants. In fact, practical implementations of branch-andbound algorithms for solving MILP problems usually determine the solution only to within a margin of optimality II E [0,1)such that objective function obtained

L

(1- A) x true objective function

Typically 1 is in the range 0.01-0.1. In the context of the modified branch-and-bound algorithm discussed here, a node will be considered to be fathomed if the corresponding objective function of P2 &e., the upper bound on the objective function of P1) does not exceed

(pLB 1-1

where q5LB is the objective function of the best feasible

solution t o P1 (Le.,the best lower bound t o the optimal solution of PI) obtained so far in the search. The modified branch-and-bound algorithm is stated formally in Figure 14. Note that with a nonzero margin of optimality A, an integer-feasible node may be fathomed even if its objective function value actually exceeds (slightly) the best lower bound (PLB found so far. In this case, we save the solution of the NLP at step 4.2.1.2, which could indeed have resulted in an improvement in @LB at step 4.2.1.3.1. On the other hand, irrespective of whether the fathomed node was integer-feasible or not, qFdeis a valid upper bound on the optimal objective function of P1, and therefore step 4.3.1 is needed to ensure that this bound is updated. However, we note that this step never results in a change in the value of the upper bound q5m if A is set to zero. It should be noted that usually the integer-feasible solution examined at step 4.2.1 results from only a partial assignment of the binary variables Wgt,with all other Wgt variables simply happening t o attain binary values in the solution of the LP relaxations (steps 4.2.2.2 and 4.2.2.3). However the NLP solved at step 4.2.1.2 is derived from P1 by fixing all Wit to their current values. It is therefore possible that an NLP corresponding to a different all-integer descendant of this node will actually have a better objective function value than the one obtained at step 4.2.1.2. This eventuality is catered for by step 4.2.1.4: the node is considered to be fathomed only if the solution of the current NLP is the best possible (Le., PLp = podel. Otherwise an integer cut excluding the current combination of Wgt variables is added to the formulation, and the corresponding linear relaxation is solved. The node is then placed back on the stack t o ensure that its integer feasible descendants are considered later in the search. Upon termination of the algorithm, we obtain a range within which the optimal objective function of P1, q5* is guaranteed to lie q5LB 5

+5*

5 +5UB

(19)

A feasible solution of P1 corresponding to q5LB is also obtained provided that q5LB > --oo. This will always be the case if at least one of the NLPs solved at step 4.2.1.2 manages to determine a feasible local minimum solution. Failures may occur at step 4.2.1.2 either because the NLP is indeed infeasible for the given values of the binary variables, or because the NLP code simply cannot obtain a solution to the problem posed to it. In practice, the probability of the latter mode of failure in the NLP code and the computational cost of solving the NLP are both drastically reduced by initializing the NLP iterations at the values of the variables obtained by solving the linear programming relaxation for this node (steps 4.2.2.2 and 4.2.2.3). The algorithm presented here bears some similarities to the LP/NLP algorithm of Quesada and Grossmann (1992). Both algorithms construct MILP approximations of MINLP problems, which they proceed to solve using branch-and-bound type search procedures. Both also solve the restricted NLPs at each integer-feasible node encountered during the search. The main difference is that the Quesada and Grossmann algorithm is restricted to the solution of conuex MINLPs and can therefore derive the approximating MILP by replacing the nonlinear functions in the original problem by Taylor expansions at a suitable initial point. In our case, because of the nonconvexity of the problem, we

Ind. Eng. Chem. Res., Vol. 33, No. 12, 1994 3181 Given a margin of optimality AE[O,l), 1

Set

lower

and

upper

p:=-,,p:=-,.

bounds

on

the

objective

function,

2

Solve linear prograrme obtained from [PZ] by relaxing the integrality requirements on the equipment allocation variables W u t .

3

If feasible, place the solution obtained on a stack containing unexplored search tree nodes.

4

WHILE the stack is not empty DO 4.1 Select and remove a solution from the stack. Denote its objective function value by $node. 4.2 IF (1 -a) pde > THEN 4.2.1 IF an integer-feasible node, THEN 4.2.1.1 Set = max qjmde). 4.2.1.2 Solve NLP problem derived from [Pll by fixing the Writ at their current integer values. If the problem is found to be feasible, then denote the objective funcbNLP, otherwise set tion value by p p : = -00. 4.2.1.3 IF # N L p 2 @ LTHEN B 4.2.1.3.1 Set # L B : = @ N L p . 4.2.1.3.2 Record values of all variables as best solution currently available. 4.2.1.4 IF ONLp 0) and endothermic (4, < 0) heat-integrated

operations. Case I: Exothermic Heat-IntegratedOperations. In this case, q v > 0 and T, > It can be seen from (A.l) that f, is maximized if the quantity (T,- Osu)/qw is minimized. This is achieved when 0,is a t its upper bound. Since the other two variables, and q,, vary over the duration of the task, we must select the time at which this quantity is minimized. Furthermore we note that the heat load qv(t)is, in fact, a non-decreasing function of batch size, and in order to minimize (T,- Osu)/qvwe have to consider the heat load for the largest possible batch size of task i taking place in unit j . This is denoted by The above reasoning, together with the pumping rate constraint (16), leads to the bound

The bounds on the values of fi(t) and Tc(t)at the end of each discrete interval-are similar to (A.5) and-(A.6) with the difference that Tiis must be replaced by Tij,e+l:

-

.

(A.7)

c".

From (A.2) we see that in order to maximize @yt, Os, must attain its minimum value and the quantity (Tu- qg/U&j) its maximum. Furthermore, we must consider the batch size that minimizes qij, i.e., the minimum batch size for task i taking place in unit j , denoted by We therefore have

Case 11: Endothermic Heat-IntegratedOperations. In this case q" < 0 and Tu < @". To facilitate the derivation of the bounds, we write (A.l) and (A.2) in the form

c".

and With similar reasoning, we obtain the following lower bound on Ort. We can now see from (A.1') that f i j is maximized if the - Tw)/(-q~) is minimized. This is achieved quantity (Os" when 0," is at its lower bound and the batch size is set a t its maximum value. This leads to

3186 Ind. Eng. Chem. Res., Vol. 33, No. 12,1994

Literature Cited

From (A.2') we see that in order to maximize Or:, 0," must attain its minimum value and -qij its

maxi-

mum. We therefore have

On the other hand, O r t is minimized by setting 0," t o its maximum value and the batch size to its minimum:

Just as for case I, we now proceed to derive the following bounds on the variables in the discrete time domain:

r

1

1

I

(A.11)

(A.14a)

1 y!? we + y!?'V""I w ui/ij

I-

Corominas, J.; Espuiia, A,; Puigjaner, L. A New Look at Energy Integration in Multiproduct Batch Processes. Comput. Chem. E71.g. 1993,17S,S15-S20. Drud, A. S. CONOPT-A GRG code for Large-Scale Nonlinear Optimization-Reference Manual; ARKI Consulting and Development A/S: Bagsvaerd, Denmark, 1992. Ivanov, B.; Peneva, K.; Bancheva, N. Heat Integration of Batch Vessels at Fixed Time Interval. I. Schemes with Recycling Main Fluids. Hung. J. Znd. Chem. 1992,20,225-231. Kemp, I. C.; MacDonald, E. K. Energy and Process Integration in Continuous and Batch Processes. Znst. Chem. Eng. Symp. Ser. 1987,105,185-200. Kemp, I. C.; MacDonald, E. K. Application of Pinch Technology to Separation, Reaction and Batch Processes. Znst. Chem. Eng. Symp. Ser. 1988,109,239-257. Kemp, I. C.; Deakin, A. W. The Cascade Analysis for Energy and Process Integration of Batch Processes. Part 1: Calculation of Energy Targets. Chem. Eng. Res. Des. 1989,67,495-509. Kondili, E.; Pantelides, C. C.; Sargent, R. W. H. A General Algorithm for Short-Term Scheduling of Batch Operations. Part I-Mathematical Formulation. Comput. Chem. Eng. 1993,17, 211-227. Linnhoff, B.; Townsend, D. W.; Boland, D.; Hewitt, G. F.; Thomas, B. E. A.; Guy, A. R.; Marsland, R. H. User Guide on Process Integration for the E m i e n t Use ofEnergy;IChemE: Rugby, UK, 1982. Linnhoff, B.; Ashton, G. J.; Obeng, E. D. A. Process Integration of Batch Processes. Znst. Chem. Eng. Symp. Ser. 1988,109,221237. Lovell-Smith, J. E. R.; Baldwin, A. J. Energy Saving Opportunities in Dairy Manufacturing Plant. N . 2.J. Dairy Sei. Technol. 1988,23,61-74. McCormick, G. P. Computability of Global Solutions to Factorable Nonconvex Programs: Part I-Convex Underestimating Problems. Math. Program. 1976,10, 147-175. Mignon, D.; Hernia, J. Using BATCHES for Modeling and Optimizing the Brewhouses of an Industrial Brewery. Comput. Chem. Eng. 1993,17S,S51-S56. Murtagh, B. A.; Saunders, M. A. MZNOS 5.0 User's Guide. Technical Report SOL 83-20,Stanford University, 1983. Obeng, E. D. A.; Ashton, G. J. On Pinch Technology Based Procedures for the Design of Batch Processes. Chem. Eng. Res. Des. 1988,66, 255-259. Pantelides, C. C. Unified Frameworks for Optimal Process Planning and Scheduling. Proceedings of the Second Conference on Foundations of Computer-Aided Operations; Rippin, D. W. T., Hale, J., Eds.; CACHE Publications: 1994. Papageorgiou, L. G.; Charalambides, M. S.; Shah, N.; Pantelides, C. C. Optimal Operation of Thermally Coupled Batch Processes. Proc. ESCAPE'4 Conf.; IChemE: Rugby, UK, 1994; pp 71-78. Peneva, K.; Ivanov, B.; Bancheva, N. Heat Integration of Batch Vessels at Fixed Time Interval. 11. Schemes with Intermediate Heating and Cooling Agents. Hung. J. Znd. Chem. 1992,20, 233-239. Quesada, I.; Grossmann, I. E. An LP/NLP Based Branch and Bound Algorithm for Convex MINLP Optimization Problems. Comput. Chem. Eng. 1992,16,937-947. Vaselenak, J. A.; Grossmann, I. E.; Westerberg, A. W. Heat Integration in Batch Processing. Znd. Eng. Chem. Process Des. Dev. 1986,25,357-366.

(A.14b)

Received for review J u n e 15, 1994 Revised manuscript received August 22, 1994*

The bounds (A.5)-(A.8) and (A.ll)-(A.l4) are those used in the linearizations (18) of the bilinear terms defined by (17).

Abstract published in Advance A C S Abstracts, October 15, 1994.

0;"

@